DynamicLens: A Dynamic User-Interface for a Meta-Recommendation System J. Ben Schafer Department of Computer Science University of Northern Iowa Cedar Falls, IA 50614-0507 USA +1 319 273 2187 [email protected] ABSTRACT Recommender Systems have emerged as powerful tools for Recommendation systems help users find items of interest. helping users reduce information overload. Such systems Meta-recommendation systems provide users with employ a variety of techniques to help users identify items personalized control over the combination of of interest [5]. For example, a recommender system in the recommendation data from multiple information sources. domain of movies might suggest that a user go see Ladder In the process, they provide users with more helpful 49 because she requested films classified as “drama” recommendations by allowing users to indicate how (query fit using information retrieval), because she has important each parameter is in their decision process, and previously liked films starring John Travolta how data should be weighted during recommendation (personalization using information filtering), or because generation. Most current meta-recommendation systems people like her have indicated it was a movie they enjoyed require the submission of large, form-based queries prior to (personalization using collaborative filtering). Regardless the receipt of recommendations. Such systems make it of technique, these systems attempt to help users identify difficult for a user to conclude what effect a given the items that best fit their needs, their tastes, or even both. requirement has on the overall recommendations. This This paper discusses a class of recommendation interface paper considers the construction of an interface to allow known as meta-recommendation systems. These systems dynamic queries for a meta-recommender. It is believed present recommendations fused from "recommendation that the addition of a dynamic query interface will provide data" from multiple information sources. Meta- users with more meaningful meta-recommendations by recommendations systems encourage users to provide both allowing them to explore these causes and effects. ephemeral and persistent information requirements. The Keywords systems use this data to produce recommendations that Meta-recommendation systems recommendation system, blend query-fit with long-term personalization. For collaborative filtering, dynamic query interface. example, MetaLens – a real- time meta-recommender in the domain of movies – might recommend Ladder 49 based not INTRODUCTION only on a combination of the reasons previously On a daily basis we are faced with information overload as mentioned, but also based on the fact that the user indicated we choose from an overwhelming number of options. To she requires a movie showing in her local theater that starts keep abreast of the latest developments in our career field, after 9:00 PM. Furthermore, these systems provide a high we can choose from a whole host of journal articles, level of user control over the combination of conference proceedings, textbooks, and web sites. During recommendation data, providing users with more unified our personal time we must choose which television show to and meaningful recommendations. This paper also watch, which movie to see, which CD to play, or which presents early work in the development of a dynamic query book to read. The number of options from which to choose interface for a meta-recommendation system. in each of these categories is often more than we can possibly process. RELATED WORK The earliest "recommender systems" were information filtering and retrieval systems designed to fight information overload in textual domains. Recommender systems that Copyright is held by the author/owner(s). incorporate information retrieval methods are frequently Workshop: Beyond Personalization 2005 used to satisfy ephemeral information needs from relatively IUI'05, January 9, 2005, San Diego, California, USA static databases. Conversely, recommender systems that http://www.cs.umn.edu/Research/GroupLens/beyond2005 incorporate information filtering (IF) methods are shoppers creating shopping lists on a personal digital frequently used to identify items that match relatively assistant (PDA). The SmartPad system considers a stable and specific information needs in domains with a consumer’s purchases across a store’s product taxonomy. rapid turnover or frequent additions. Although information Recommendations of product subclasses are based upon a retrieval and information filtering are considered combination of class and subclass associations drawn from fundamentally different tasks [1], they are based on similar information filtering and co-purchase rules drawn from techniques. This paper will consider both under the data mining. Product rankings within a product subclass singular term "information filtering." Systems which use are based upon the products’ sales rankings within the IF techniques include RE:Agent [2], Ripper [5], NewT [9], user’s consumer cluster, a less personalized variation of and Amalthaea [10]. collaborative filtering. Collaborative filtering (CF) is an attempt to facilitate the Nakamura and Abe [11] describe a system for the process of "word of mouth." Users provide the system automatic recording of programs using a digital video with evaluations of items that may be used to make recorder. They implement a set of “specialist” algorithms "recommendations" to other users. The simplest of CF that use probabilistic estimation to produce systems provides generalized recommendations by recommendations that are both content-based (based on aggregating the evaluations of the community at large. information about previously recorded shows from the More advanced systems personalize the process by forming electronic program guide) and collaborative (based on the an individualized neighborhood for each user consisting of viewing patterns of similar users). a subset of users whose opinions are highly correlated with META-RECOMMENDERS those of the original user. Recommender systems based on Consider the following scenario. Mary's 8-year-old collaborative filtering include MovieLens [6], Tapestry nephew is visiting for the weekend, and she would like to [7], GroupLens [12], Ringo [17], and PHOAKS [19]. take him to the movies. Mary has several criteria for the Hybrid Recommender Systems movie that she will select. She would like a comedy or As researchers have studied different recommender system family movie rated no "higher" than PG-13. She would technologies, many have suggested that no single prefer that the movie contain no sex, violence or offensive technology works for all situations. Thus, hybrid systems language, last less than two hours and, if possible, show at have been built in an attempt to use the strengths of one a theater in her neighborhood. Finally, she would like to technology to offset the weaknesses of another. Burke [3] select a movie that she herself might enjoy. discusses several different hybridization methods, but Traditionally, Mary might decide which movie to see by points out that most hybrid systems involve the checking the theater listings in the newspaper and asking combination of collaborative filtering with either a content- friends for recommendations. More recently, her quest based (IF) or data mining technique. might include the use of the Internet to access online Tango [4] recommends articles in the domain of an online theater listings and search databases of movie reviews. newspaper. It does so by creating separate Additionally, she might be able to obtain personalized, CF- recommendations from CF and IF algorithms and merging based recommendations from a web site such as these using a separate combination filter. The combination MovieLens. Producing her final selection, however, filter employed by Tango uses per-user, per-article weights. requires a significant amount of manual intervention; Mary The calculation of these weights takes into account the must visit each source to gather the data and then decide degree of confidence each filter has in a particular how to apply this data in making her final decision. document’s recommendation, as well as error analysis for The hybrid systems mentioned in the previous section are a each filter’s past performance for the user in question. significant step toward solving problems like Mary’s. A Torres et al. [20] present the results of several experiments hybrid movie recommendation system would provide Mary involving TechLens. Similar to Tango, TechLens with lists of movies blended from her long-standing combines both a collaborative filter and a content-based collaborative filtering and content-interest profiles. It is filter to recommend research papers. In both offline and likely, however, that such a system would not offer her the online studies they consider five different algorithms for ability to provide information that might improve the combining the recommendations from these filters, recommendations produced by the combination algorithm. including sequential algorithms. These techniques take the For example, if given access to the combination algorithm, recommendations from one filter as a seed to the second Mary could indicate that predictions should be biased less filter. They conclude that different algorithms should be towards the British art films she frequently likes and more used for recommending different kinds of papers, although toward the family movies appropriate for her